Awesome
<img src="https://github.com/DeVriesMatt/cellshape-voxel/blob/main/img/cellshape_voxel.png" alt="Cellshape logo by Matt De Vries">
Cellshape-voxel is an easy-to-use tool to analyse the shapes of cells using deep learning and, in particular, 3D convolutional neural networks. The tool provides the ability to train 3D convolutional autoencoders on 3D single cell masks as well as providing pre-trained networks for inference.
To install
pip install cellshape-voxel
Usage
Basic usage
import torch
from cellshape_voxel import VoxelAutoEncoder
from cellshape_voxel.encoders.resnet import Bottleneck
model = VoxelAutoEncoder(num_layers_encoder=3,
num_layers_decoder=3,
encoder_type="resnet",
input_shape=(64, 64, 64, 1),
filters=(32, 64, 128, 256, 512),
num_features=50,
bias=True,
activations=False,
batch_norm=True,
leaky=True,
neg_slope=0.01,
resnet_depth=10,
resnet_block_inplanes=(64, 128, 256, 512),
resnet_block=Bottleneck,
n_input_channels=1,
no_max_pool=True,
resnet_shortcut_type="B",
resnet_widen_factor=1.0)
volume = torch.randn(1, 64, 64, 64, 1)
recon, features = model(volume)
To train a 3D resnet autoencoder on masks of cells or nuclei:
import torch
from torch.utils.data import DataLoader
import cellshape_voxel as voxel
input_dir = "path/to/binary/mask/files/"
batch_size = 16
learning_rate = 0.0001
num_epochs = 1
output_dir = "path/to/save/output/"
model = voxel.AutoEncoder(
num_layers_encoder=4,
num_layers_decoder=4,
input_shape=(64, 64, 64, 1),
encoder_type="resnet",
)
dataset = voxel.VoxelDataset(
PATH_TO_DATASET, transform=None, img_size=(300, 300, 300)
)
dataloader = voxel.DataLoader(dataset, batch_size=batch_size, shuffle=True)
optimizer = torch.optim.Adam(
model.parameters(),
lr=learning_rate * 16 / batch_size,
betas=(0.9, 0.999),
weight_decay=1e-6,
)
voxel.train(model, dataloader, 1, optimizer, output_dir)
Parameters
num_features
: int.
The size of the latent space of the autoencoder. If you have rectangular images, make sure your image size is the maximum of the width and heightk
: int.
The number of neightbours to use in the k-nearest-neighbours graph construction.encoder_type
: int.
The type of encoder: 'foldingnet' or 'dgcnn'decoder_type
: int.
The type of decoder: 'foldingnet' or 'dgcnn'